crypto/modes/asm/ghash-x86_64.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2010-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the Apache License 2.0 (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9 #
  10 # ====================================================================
  11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  12 # project. The module is, however, dual licensed under OpenSSL and
  13 # CRYPTOGAMS licenses depending on where you obtain it. For further
  14 # details see http://www.openssl.org/~appro/cryptogams/.
  15 # ====================================================================
  16 #
  17 # March, June 2010
  18 #
  19 # The module implements "4-bit" GCM GHASH function and underlying
  20 # single multiplication operation in GF(2^128). "4-bit" means that
  21 # it uses 256 bytes per-key table [+128 bytes shared table]. GHASH
  22 # function features so called "528B" variant utilizing additional
  23 # 256+16 bytes of per-key storage [+512 bytes shared table].
  24 # Performance results are for this streamed GHASH subroutine and are
  25 # expressed in cycles per processed byte, less is better:
  26 #
  27 #               gcc 3.4.x(*)    assembler
  28 #
  29 # P4            28.6            14.0            +100%
  30 # Opteron       19.3            7.7             +150%
  31 # Core2         17.8            8.1(**)         +120%
  32 # Atom          31.6            16.8            +88%
  33 # VIA Nano      21.8            10.1            +115%
  34 #
  35 # (*)   comparison is not completely fair, because C results are
  36 #       for vanilla "256B" implementation, while assembler results
  37 #       are for "528B";-)
  38 # (**)  it's mystery [to me] why Core2 result is not same as for
  39 #       Opteron;
  40
  41 # May 2010
  42 #
  43 # Add PCLMULQDQ version performing at 2.02 cycles per processed byte.
  44 # See ghash-x86.pl for background information and details about coding
  45 # techniques.
  46 #
  47 # Special thanks to David Woodhouse for providing access to a
  48 # Westmere-based system on behalf of Intel Open Source Technology Centre.
  49
  50 # December 2012
  51 #
  52 # Overhaul: aggregate Karatsuba post-processing, improve ILP in
  53 # reduction_alg9, increase reduction aggregate factor to 4x. As for
  54 # the latter. ghash-x86.pl discusses that it makes lesser sense to
  55 # increase aggregate factor. Then why increase here? Critical path
  56 # consists of 3 independent pclmulqdq instructions, Karatsuba post-
  57 # processing and reduction. "On top" of this we lay down aggregated
  58 # multiplication operations, triplets of independent pclmulqdq's. As
  59 # issue rate for pclmulqdq is limited, it makes lesser sense to
  60 # aggregate more multiplications than it takes to perform remaining
  61 # non-multiplication operations. 2x is near-optimal coefficient for
  62 # contemporary Intel CPUs (therefore modest improvement coefficient),
  63 # but not for Bulldozer. Latter is because logical SIMD operations
  64 # are twice as slow in comparison to Intel, so that critical path is
  65 # longer. A CPU with higher pclmulqdq issue rate would also benefit
  66 # from higher aggregate factor...
  67 #
  68 # Westmere      1.78(+13%)
  69 # Sandy Bridge  1.80(+8%)
  70 # Ivy Bridge    1.80(+7%)
  71 # Haswell       0.55(+93%) (if system doesn't support AVX)
  72 # Broadwell     0.45(+110%)(if system doesn't support AVX)
  73 # Skylake       0.44(+110%)(if system doesn't support AVX)
  74 # Bulldozer     1.49(+27%)
  75 # Silvermont    2.88(+13%)
  76 # Knights L     2.12(-)    (if system doesn't support AVX)
  77 # Goldmont      1.08(+24%)
  78
  79 # March 2013
  80 #
  81 # ... 8x aggregate factor AVX code path is using reduction algorithm
  82 # suggested by Shay Gueron[1]. Even though contemporary AVX-capable
  83 # CPUs such as Sandy and Ivy Bridge can execute it, the code performs
  84 # sub-optimally in comparison to above mentioned version. But thanks
  85 # to Ilya Albrekht and Max Locktyukhin of Intel Corp. we knew that
  86 # it performs in 0.41 cycles per byte on Haswell processor, in
  87 # 0.29 on Broadwell, and in 0.36 on Skylake.
  88 #
  89 # Knights Landing achieves 1.09 cpb.
  90 #
  91 # [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
  92
  93 $flavour = shift;
  94 $output  = shift;
  95 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  96
  97 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  98
  99 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 100 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
 101 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 102 die "can't locate x86_64-xlate.pl";
 103
 104 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
 105                 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
 106         $avx = ($1>=2.20) + ($1>=2.22);
 107 }
 108
 109 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
 110             `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
 111         $avx = ($1>=2.09) + ($1>=2.10);
 112 }
 113
 114 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
 115             `ml64 2>&1` =~ /Version ([0-9]+)\./) {
 116         $avx = ($1>=10) + ($1>=11);
 117 }
 118
 119 if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9]\.[0-9]+)/) {
 120         $avx = ($2>=3.0) + ($2>3.0);
 121 }
 122
 123 open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
 124 *STDOUT=*OUT;
 125
 126 $do4xaggr=1;
 127
 128 # common register layout
 129 $nlo="%rax";
 130 $nhi="%rbx";
 131 $Zlo="%r8";
 132 $Zhi="%r9";
 133 $tmp="%r10";
 134 $rem_4bit = "%r11";
 135
 136 $Xi="%rdi";
 137 $Htbl="%rsi";
 138
 139 # per-function register layout
 140 $cnt="%rcx";
 141 $rem="%rdx";
 142
 143 sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/     or
 144                         $r =~ s/%[er]([sd]i)/%\1l/      or
 145                         $r =~ s/%[er](bp)/%\1l/         or
 146                         $r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
 147
 148 sub AUTOLOAD()          # thunk [simplified] 32-bit style perlasm
 149 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
 150   my $arg = pop;
 151     $arg = "\$$arg" if ($arg*1 eq $arg);
 152     $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
 153 }
 154 \f
 155 { my $N;
 156   sub loop() {
 157   my $inp = shift;
 158
 159         $N++;
 160 $code.=<<___;
 161         xor     $nlo,$nlo
 162         xor     $nhi,$nhi
 163         mov     `&LB("$Zlo")`,`&LB("$nlo")`
 164         mov     `&LB("$Zlo")`,`&LB("$nhi")`
 165         shl     \$4,`&LB("$nlo")`
 166         mov     \$14,$cnt
 167         mov     8($Htbl,$nlo),$Zlo
 168         mov     ($Htbl,$nlo),$Zhi
 169         and     \$0xf0,`&LB("$nhi")`
 170         mov     $Zlo,$rem
 171         jmp     .Loop$N
 172
 173 .align  16
 174 .Loop$N:
 175         shr     \$4,$Zlo
 176         and     \$0xf,$rem
 177         mov     $Zhi,$tmp
 178         mov     ($inp,$cnt),`&LB("$nlo")`
 179         shr     \$4,$Zhi
 180         xor     8($Htbl,$nhi),$Zlo
 181         shl     \$60,$tmp
 182         xor     ($Htbl,$nhi),$Zhi
 183         mov     `&LB("$nlo")`,`&LB("$nhi")`
 184         xor     ($rem_4bit,$rem,8),$Zhi
 185         mov     $Zlo,$rem
 186         shl     \$4,`&LB("$nlo")`
 187         xor     $tmp,$Zlo
 188         dec     $cnt
 189         js      .Lbreak$N
 190
 191         shr     \$4,$Zlo
 192         and     \$0xf,$rem
 193         mov     $Zhi,$tmp
 194         shr     \$4,$Zhi
 195         xor     8($Htbl,$nlo),$Zlo
 196         shl     \$60,$tmp
 197         xor     ($Htbl,$nlo),$Zhi
 198         and     \$0xf0,`&LB("$nhi")`
 199         xor     ($rem_4bit,$rem,8),$Zhi
 200         mov     $Zlo,$rem
 201         xor     $tmp,$Zlo
 202         jmp     .Loop$N
 203
 204 .align  16
 205 .Lbreak$N:
 206         shr     \$4,$Zlo
 207         and     \$0xf,$rem
 208         mov     $Zhi,$tmp
 209         shr     \$4,$Zhi
 210         xor     8($Htbl,$nlo),$Zlo
 211         shl     \$60,$tmp
 212         xor     ($Htbl,$nlo),$Zhi
 213         and     \$0xf0,`&LB("$nhi")`
 214         xor     ($rem_4bit,$rem,8),$Zhi
 215         mov     $Zlo,$rem
 216         xor     $tmp,$Zlo
 217
 218         shr     \$4,$Zlo
 219         and     \$0xf,$rem
 220         mov     $Zhi,$tmp
 221         shr     \$4,$Zhi
 222         xor     8($Htbl,$nhi),$Zlo
 223         shl     \$60,$tmp
 224         xor     ($Htbl,$nhi),$Zhi
 225         xor     $tmp,$Zlo
 226         xor     ($rem_4bit,$rem,8),$Zhi
 227
 228         bswap   $Zlo
 229         bswap   $Zhi
 230 ___
 231 }}
 232
 233 $code=<<___;
 234 .text
 235 .extern OPENSSL_ia32cap_P
 236
 237 .globl  gcm_gmult_4bit
 238 .type   gcm_gmult_4bit,\@function,2
 239 .align  16
 240 gcm_gmult_4bit:
 241 .cfi_startproc
 242         push    %rbx
 243 .cfi_push       %rbx
 244         push    %rbp            # %rbp and others are pushed exclusively in
 245 .cfi_push       %rbp
 246         push    %r12            # order to reuse Win64 exception handler...
 247 .cfi_push       %r12
 248         push    %r13
 249 .cfi_push       %r13
 250         push    %r14
 251 .cfi_push       %r14
 252         push    %r15
 253 .cfi_push       %r15
 254         sub     \$280,%rsp
 255 .cfi_adjust_cfa_offset  280
 256 .Lgmult_prologue:
 257
 258         movzb   15($Xi),$Zlo
 259         lea     .Lrem_4bit(%rip),$rem_4bit
 260 ___
 261         &loop   ($Xi);
 262 $code.=<<___;
 263         mov     $Zlo,8($Xi)
 264         mov     $Zhi,($Xi)
 265
 266         lea     280+48(%rsp),%rsi
 267 .cfi_def_cfa    %rsi,8
 268         mov     -8(%rsi),%rbx
 269 .cfi_restore    %rbx
 270         lea     (%rsi),%rsp
 271 .cfi_def_cfa_register   %rsp
 272 .Lgmult_epilogue:
 273         ret
 274 .cfi_endproc
 275 .size   gcm_gmult_4bit,.-gcm_gmult_4bit
 276 ___
 277 \f
 278 # per-function register layout
 279 $inp="%rdx";
 280 $len="%rcx";
 281 $rem_8bit=$rem_4bit;
 282
 283 $code.=<<___;
 284 .globl  gcm_ghash_4bit
 285 .type   gcm_ghash_4bit,\@function,4
 286 .align  16
 287 gcm_ghash_4bit:
 288 .cfi_startproc
 289         push    %rbx
 290 .cfi_push       %rbx
 291         push    %rbp
 292 .cfi_push       %rbp
 293         push    %r12
 294 .cfi_push       %r12
 295         push    %r13
 296 .cfi_push       %r13
 297         push    %r14
 298 .cfi_push       %r14
 299         push    %r15
 300 .cfi_push       %r15
 301         sub     \$280,%rsp
 302 .cfi_adjust_cfa_offset  280
 303 .Lghash_prologue:
 304         mov     $inp,%r14               # reassign couple of args
 305         mov     $len,%r15
 306 ___
 307 { my $inp="%r14";
 308   my $dat="%edx";
 309   my $len="%r15";
 310   my @nhi=("%ebx","%ecx");
 311   my @rem=("%r12","%r13");
 312   my $Hshr4="%rbp";
 313
 314         &sub    ($Htbl,-128);           # size optimization
 315         &lea    ($Hshr4,"16+128(%rsp)");
 316         { my @lo =($nlo,$nhi);
 317           my @hi =($Zlo,$Zhi);
 318
 319           &xor  ($dat,$dat);
 320           for ($i=0,$j=-2;$i<18;$i++,$j++) {
 321             &mov        ("$j(%rsp)",&LB($dat))          if ($i>1);
 322             &or         ($lo[0],$tmp)                   if ($i>1);
 323             &mov        (&LB($dat),&LB($lo[1]))         if ($i>0 && $i<17);
 324             &shr        ($lo[1],4)                      if ($i>0 && $i<17);
 325             &mov        ($tmp,$hi[1])                   if ($i>0 && $i<17);
 326             &shr        ($hi[1],4)                      if ($i>0 && $i<17);
 327             &mov        ("8*$j($Hshr4)",$hi[0])         if ($i>1);
 328             &mov        ($hi[0],"16*$i+0-128($Htbl)")   if ($i<16);
 329             &shl        (&LB($dat),4)                   if ($i>0 && $i<17);
 330             &mov        ("8*$j-128($Hshr4)",$lo[0])     if ($i>1);
 331             &mov        ($lo[0],"16*$i+8-128($Htbl)")   if ($i<16);
 332             &shl        ($tmp,60)                       if ($i>0 && $i<17);
 333
 334             push        (@lo,shift(@lo));
 335             push        (@hi,shift(@hi));
 336           }
 337         }
 338         &add    ($Htbl,-128);
 339         &mov    ($Zlo,"8($Xi)");
 340         &mov    ($Zhi,"0($Xi)");
 341         &add    ($len,$inp);            # pointer to the end of data
 342         &lea    ($rem_8bit,".Lrem_8bit(%rip)");
 343         &jmp    (".Louter_loop");
 344
 345 $code.=".align  16\n.Louter_loop:\n";
 346         &xor    ($Zhi,"($inp)");
 347         &mov    ("%rdx","8($inp)");
 348         &lea    ($inp,"16($inp)");
 349         &xor    ("%rdx",$Zlo);
 350         &mov    ("($Xi)",$Zhi);
 351         &mov    ("8($Xi)","%rdx");
 352         &shr    ("%rdx",32);
 353
 354         &xor    ($nlo,$nlo);
 355         &rol    ($dat,8);
 356         &mov    (&LB($nlo),&LB($dat));
 357         &movz   ($nhi[0],&LB($dat));
 358         &shl    (&LB($nlo),4);
 359         &shr    ($nhi[0],4);
 360
 361         for ($j=11,$i=0;$i<15;$i++) {
 362             &rol        ($dat,8);
 363             &xor        ($Zlo,"8($Htbl,$nlo)")                  if ($i>0);
 364             &xor        ($Zhi,"($Htbl,$nlo)")                   if ($i>0);
 365             &mov        ($Zlo,"8($Htbl,$nlo)")                  if ($i==0);
 366             &mov        ($Zhi,"($Htbl,$nlo)")                   if ($i==0);
 367
 368             &mov        (&LB($nlo),&LB($dat));
 369             &xor        ($Zlo,$tmp)                             if ($i>0);
 370             &movzw      ($rem[1],"($rem_8bit,$rem[1],2)")       if ($i>0);
 371
 372             &movz       ($nhi[1],&LB($dat));
 373             &shl        (&LB($nlo),4);
 374             &movzb      ($rem[0],"(%rsp,$nhi[0])");
 375
 376             &shr        ($nhi[1],4)                             if ($i<14);
 377             &and        ($nhi[1],0xf0)                          if ($i==14);
 378             &shl        ($rem[1],48)                            if ($i>0);
 379             &xor        ($rem[0],$Zlo);
 380
 381             &mov        ($tmp,$Zhi);
 382             &xor        ($Zhi,$rem[1])                          if ($i>0);
 383             &shr        ($Zlo,8);
 384
 385             &movz       ($rem[0],&LB($rem[0]));
 386             &mov        ($dat,"$j($Xi)")                        if (--$j%4==0);
 387             &shr        ($Zhi,8);
 388
 389             &xor        ($Zlo,"-128($Hshr4,$nhi[0],8)");
 390             &shl        ($tmp,56);
 391             &xor        ($Zhi,"($Hshr4,$nhi[0],8)");
 392
 393             unshift     (@nhi,pop(@nhi));               # "rotate" registers
 394             unshift     (@rem,pop(@rem));
 395         }
 396         &movzw  ($rem[1],"($rem_8bit,$rem[1],2)");
 397         &xor    ($Zlo,"8($Htbl,$nlo)");
 398         &xor    ($Zhi,"($Htbl,$nlo)");
 399
 400         &shl    ($rem[1],48);
 401         &xor    ($Zlo,$tmp);
 402
 403         &xor    ($Zhi,$rem[1]);
 404         &movz   ($rem[0],&LB($Zlo));
 405         &shr    ($Zlo,4);
 406
 407         &mov    ($tmp,$Zhi);
 408         &shl    (&LB($rem[0]),4);
 409         &shr    ($Zhi,4);
 410
 411         &xor    ($Zlo,"8($Htbl,$nhi[0])");
 412         &movzw  ($rem[0],"($rem_8bit,$rem[0],2)");
 413         &shl    ($tmp,60);
 414
 415         &xor    ($Zhi,"($Htbl,$nhi[0])");
 416         &xor    ($Zlo,$tmp);
 417         &shl    ($rem[0],48);
 418
 419         &bswap  ($Zlo);
 420         &xor    ($Zhi,$rem[0]);
 421
 422         &bswap  ($Zhi);
 423         &cmp    ($inp,$len);
 424         &jb     (".Louter_loop");
 425 }
 426 $code.=<<___;
 427         mov     $Zlo,8($Xi)
 428         mov     $Zhi,($Xi)
 429
 430         lea     280+48(%rsp),%rsi
 431 .cfi_def_cfa    %rsi,8
 432         mov     -48(%rsi),%r15
 433 .cfi_restore    %r15
 434         mov     -40(%rsi),%r14
 435 .cfi_restore    %r14
 436         mov     -32(%rsi),%r13
 437 .cfi_restore    %r13
 438         mov     -24(%rsi),%r12
 439 .cfi_restore    %r12
 440         mov     -16(%rsi),%rbp
 441 .cfi_restore    %rbp
 442         mov     -8(%rsi),%rbx
 443 .cfi_restore    %rbx
 444         lea     0(%rsi),%rsp
 445 .cfi_def_cfa_register   %rsp
 446 .Lghash_epilogue:
 447         ret
 448 .cfi_endproc
 449 .size   gcm_ghash_4bit,.-gcm_ghash_4bit
 450 ___
 451 \f
 452 ######################################################################
 453 # PCLMULQDQ version.
 454
 455 @_4args=$win64? ("%rcx","%rdx","%r8", "%r9") :  # Win64 order
 456                 ("%rdi","%rsi","%rdx","%rcx");  # Unix order
 457
 458 ($Xi,$Xhi)=("%xmm0","%xmm1");   $Hkey="%xmm2";
 459 ($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
 460
 461 sub clmul64x64_T2 {     # minimal register pressure
 462 my ($Xhi,$Xi,$Hkey,$HK)=@_;
 463
 464 if (!defined($HK)) {    $HK = $T2;
 465 $code.=<<___;
 466         movdqa          $Xi,$Xhi                #
 467         pshufd          \$0b01001110,$Xi,$T1
 468         pshufd          \$0b01001110,$Hkey,$T2
 469         pxor            $Xi,$T1                 #
 470         pxor            $Hkey,$T2
 471 ___
 472 } else {
 473 $code.=<<___;
 474         movdqa          $Xi,$Xhi                #
 475         pshufd          \$0b01001110,$Xi,$T1
 476         pxor            $Xi,$T1                 #
 477 ___
 478 }
 479 $code.=<<___;
 480         pclmulqdq       \$0x00,$Hkey,$Xi        #######
 481         pclmulqdq       \$0x11,$Hkey,$Xhi       #######
 482         pclmulqdq       \$0x00,$HK,$T1          #######
 483         pxor            $Xi,$T1                 #
 484         pxor            $Xhi,$T1                #
 485
 486         movdqa          $T1,$T2                 #
 487         psrldq          \$8,$T1
 488         pslldq          \$8,$T2                 #
 489         pxor            $T1,$Xhi
 490         pxor            $T2,$Xi                 #
 491 ___
 492 }
 493
 494 sub reduction_alg9 {    # 17/11 times faster than Intel version
 495 my ($Xhi,$Xi) = @_;
 496
 497 $code.=<<___;
 498         # 1st phase
 499         movdqa          $Xi,$T2                 #
 500         movdqa          $Xi,$T1
 501         psllq           \$5,$Xi
 502         pxor            $Xi,$T1                 #
 503         psllq           \$1,$Xi
 504         pxor            $T1,$Xi                 #
 505         psllq           \$57,$Xi                #
 506         movdqa          $Xi,$T1                 #
 507         pslldq          \$8,$Xi
 508         psrldq          \$8,$T1                 #
 509         pxor            $T2,$Xi
 510         pxor            $T1,$Xhi                #
 511
 512         # 2nd phase
 513         movdqa          $Xi,$T2
 514         psrlq           \$1,$Xi
 515         pxor            $T2,$Xhi                #
 516         pxor            $Xi,$T2
 517         psrlq           \$5,$Xi
 518         pxor            $T2,$Xi                 #
 519         psrlq           \$1,$Xi                 #
 520         pxor            $Xhi,$Xi                #
 521 ___
 522 }
 523 \f
 524 { my ($Htbl,$Xip)=@_4args;
 525   my $HK="%xmm6";
 526
 527 $code.=<<___;
 528 .globl  gcm_init_clmul
 529 .type   gcm_init_clmul,\@abi-omnipotent
 530 .align  16
 531 gcm_init_clmul:
 532 .cfi_startproc
 533 .L_init_clmul:
 534 ___
 535 $code.=<<___ if ($win64);
 536 .LSEH_begin_gcm_init_clmul:
 537         # I can't trust assembler to use specific encoding:-(
 538         .byte   0x48,0x83,0xec,0x18             #sub    $0x18,%rsp
 539         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
 540 ___
 541 $code.=<<___;
 542         movdqu          ($Xip),$Hkey
 543         pshufd          \$0b01001110,$Hkey,$Hkey        # dword swap
 544
 545         # <<1 twist
 546         pshufd          \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
 547         movdqa          $Hkey,$T1
 548         psllq           \$1,$Hkey
 549         pxor            $T3,$T3                 #
 550         psrlq           \$63,$T1
 551         pcmpgtd         $T2,$T3                 # broadcast carry bit
 552         pslldq          \$8,$T1
 553         por             $T1,$Hkey               # H<<=1
 554
 555         # magic reduction
 556         pand            .L0x1c2_polynomial(%rip),$T3
 557         pxor            $T3,$Hkey               # if(carry) H^=0x1c2_polynomial
 558
 559         # calculate H^2
 560         pshufd          \$0b01001110,$Hkey,$HK
 561         movdqa          $Hkey,$Xi
 562         pxor            $Hkey,$HK
 563 ___
 564         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);
 565         &reduction_alg9 ($Xhi,$Xi);
 566 $code.=<<___;
 567         pshufd          \$0b01001110,$Hkey,$T1
 568         pshufd          \$0b01001110,$Xi,$T2
 569         pxor            $Hkey,$T1               # Karatsuba pre-processing
 570         movdqu          $Hkey,0x00($Htbl)       # save H
 571         pxor            $Xi,$T2                 # Karatsuba pre-processing
 572         movdqu          $Xi,0x10($Htbl)         # save H^2
 573         palignr         \$8,$T1,$T2             # low part is H.lo^H.hi...
 574         movdqu          $T2,0x20($Htbl)         # save Karatsuba "salt"
 575 ___
 576 if ($do4xaggr) {
 577         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H^3
 578         &reduction_alg9 ($Xhi,$Xi);
 579 $code.=<<___;
 580         movdqa          $Xi,$T3
 581 ___
 582         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H^4
 583         &reduction_alg9 ($Xhi,$Xi);
 584 $code.=<<___;
 585         pshufd          \$0b01001110,$T3,$T1
 586         pshufd          \$0b01001110,$Xi,$T2
 587         pxor            $T3,$T1                 # Karatsuba pre-processing
 588         movdqu          $T3,0x30($Htbl)         # save H^3
 589         pxor            $Xi,$T2                 # Karatsuba pre-processing
 590         movdqu          $Xi,0x40($Htbl)         # save H^4
 591         palignr         \$8,$T1,$T2             # low part is H^3.lo^H^3.hi...
 592         movdqu          $T2,0x50($Htbl)         # save Karatsuba "salt"
 593 ___
 594 }
 595 $code.=<<___ if ($win64);
 596         movaps  (%rsp),%xmm6
 597         lea     0x18(%rsp),%rsp
 598 .LSEH_end_gcm_init_clmul:
 599 ___
 600 $code.=<<___;
 601         ret
 602 .cfi_endproc
 603 .size   gcm_init_clmul,.-gcm_init_clmul
 604 ___
 605 }
 606
 607 { my ($Xip,$Htbl)=@_4args;
 608
 609 $code.=<<___;
 610 .globl  gcm_gmult_clmul
 611 .type   gcm_gmult_clmul,\@abi-omnipotent
 612 .align  16
 613 gcm_gmult_clmul:
 614 .cfi_startproc
 615 .L_gmult_clmul:
 616         movdqu          ($Xip),$Xi
 617         movdqa          .Lbswap_mask(%rip),$T3
 618         movdqu          ($Htbl),$Hkey
 619         movdqu          0x20($Htbl),$T2
 620         pshufb          $T3,$Xi
 621 ___
 622         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$T2);
 623 $code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
 624         # experimental alternative. special thing about is that there
 625         # no dependency between the two multiplications...
 626         mov             \$`0xE1<<1`,%eax
 627         mov             \$0xA040608020C0E000,%r10       # ((7..0)·0xE0)&0xff
 628         mov             \$0x07,%r11d
 629         movq            %rax,$T1
 630         movq            %r10,$T2
 631         movq            %r11,$T3                # borrow $T3
 632         pand            $Xi,$T3
 633         pshufb          $T3,$T2                 # ($Xi&7)·0xE0
 634         movq            %rax,$T3
 635         pclmulqdq       \$0x00,$Xi,$T1          # ·(0xE1<<1)
 636         pxor            $Xi,$T2
 637         pslldq          \$15,$T2
 638         paddd           $T2,$T2                 # <<(64+56+1)
 639         pxor            $T2,$Xi
 640         pclmulqdq       \$0x01,$T3,$Xi
 641         movdqa          .Lbswap_mask(%rip),$T3  # reload $T3
 642         psrldq          \$1,$T1
 643         pxor            $T1,$Xhi
 644         pslldq          \$7,$Xi
 645         pxor            $Xhi,$Xi
 646 ___
 647 $code.=<<___;
 648         pshufb          $T3,$Xi
 649         movdqu          $Xi,($Xip)
 650         ret
 651 .cfi_endproc
 652 .size   gcm_gmult_clmul,.-gcm_gmult_clmul
 653 ___
 654 }
 655 \f
 656 { my ($Xip,$Htbl,$inp,$len)=@_4args;
 657   my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(3..7));
 658   my ($T1,$T2,$T3)=map("%xmm$_",(8..10));
 659
 660 $code.=<<___;
 661 .globl  gcm_ghash_clmul
 662 .type   gcm_ghash_clmul,\@abi-omnipotent
 663 .align  32
 664 gcm_ghash_clmul:
 665 .cfi_startproc
 666 .L_ghash_clmul:
 667 ___
 668 $code.=<<___ if ($win64);
 669         lea     -0x88(%rsp),%rax
 670 .LSEH_begin_gcm_ghash_clmul:
 671         # I can't trust assembler to use specific encoding:-(
 672         .byte   0x48,0x8d,0x60,0xe0             #lea    -0x20(%rax),%rsp
 673         .byte   0x0f,0x29,0x70,0xe0             #movaps %xmm6,-0x20(%rax)
 674         .byte   0x0f,0x29,0x78,0xf0             #movaps %xmm7,-0x10(%rax)
 675         .byte   0x44,0x0f,0x29,0x00             #movaps %xmm8,0(%rax)
 676         .byte   0x44,0x0f,0x29,0x48,0x10        #movaps %xmm9,0x10(%rax)
 677         .byte   0x44,0x0f,0x29,0x50,0x20        #movaps %xmm10,0x20(%rax)
 678         .byte   0x44,0x0f,0x29,0x58,0x30        #movaps %xmm11,0x30(%rax)
 679         .byte   0x44,0x0f,0x29,0x60,0x40        #movaps %xmm12,0x40(%rax)
 680         .byte   0x44,0x0f,0x29,0x68,0x50        #movaps %xmm13,0x50(%rax)
 681         .byte   0x44,0x0f,0x29,0x70,0x60        #movaps %xmm14,0x60(%rax)
 682         .byte   0x44,0x0f,0x29,0x78,0x70        #movaps %xmm15,0x70(%rax)
 683 ___
 684 $code.=<<___;
 685         movdqa          .Lbswap_mask(%rip),$T3
 686
 687         movdqu          ($Xip),$Xi
 688         movdqu          ($Htbl),$Hkey
 689         movdqu          0x20($Htbl),$HK
 690         pshufb          $T3,$Xi
 691
 692         sub             \$0x10,$len
 693         jz              .Lodd_tail
 694
 695         movdqu          0x10($Htbl),$Hkey2
 696 ___
 697 if ($do4xaggr) {
 698 my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
 699
 700 $code.=<<___;
 701         mov             OPENSSL_ia32cap_P+4(%rip),%eax
 702         cmp             \$0x30,$len
 703         jb              .Lskip4x
 704
 705         and             \$`1<<26|1<<22`,%eax    # isolate MOVBE+XSAVE
 706         cmp             \$`1<<22`,%eax          # check for MOVBE without XSAVE
 707         je              .Lskip4x
 708
 709         sub             \$0x30,$len
 710         mov             \$0xA040608020C0E000,%rax       # ((7..0)·0xE0)&0xff
 711         movdqu          0x30($Htbl),$Hkey3
 712         movdqu          0x40($Htbl),$Hkey4
 713
 714         #######
 715         # Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
 716         #
 717         movdqu          0x30($inp),$Xln
 718          movdqu         0x20($inp),$Xl
 719         pshufb          $T3,$Xln
 720          pshufb         $T3,$Xl
 721         movdqa          $Xln,$Xhn
 722         pshufd          \$0b01001110,$Xln,$Xmn
 723         pxor            $Xln,$Xmn
 724         pclmulqdq       \$0x00,$Hkey,$Xln
 725         pclmulqdq       \$0x11,$Hkey,$Xhn
 726         pclmulqdq       \$0x00,$HK,$Xmn
 727
 728         movdqa          $Xl,$Xh
 729         pshufd          \$0b01001110,$Xl,$Xm
 730         pxor            $Xl,$Xm
 731         pclmulqdq       \$0x00,$Hkey2,$Xl
 732         pclmulqdq       \$0x11,$Hkey2,$Xh
 733         pclmulqdq       \$0x10,$HK,$Xm
 734         xorps           $Xl,$Xln
 735         xorps           $Xh,$Xhn
 736         movups          0x50($Htbl),$HK
 737         xorps           $Xm,$Xmn
 738
 739         movdqu          0x10($inp),$Xl
 740          movdqu         0($inp),$T1
 741         pshufb          $T3,$Xl
 742          pshufb         $T3,$T1
 743         movdqa          $Xl,$Xh
 744         pshufd          \$0b01001110,$Xl,$Xm
 745          pxor           $T1,$Xi
 746         pxor            $Xl,$Xm
 747         pclmulqdq       \$0x00,$Hkey3,$Xl
 748          movdqa         $Xi,$Xhi
 749          pshufd         \$0b01001110,$Xi,$T1
 750          pxor           $Xi,$T1
 751         pclmulqdq       \$0x11,$Hkey3,$Xh
 752         pclmulqdq       \$0x00,$HK,$Xm
 753         xorps           $Xl,$Xln
 754         xorps           $Xh,$Xhn
 755
 756         lea     0x40($inp),$inp
 757         sub     \$0x40,$len
 758         jc      .Ltail4x
 759
 760         jmp     .Lmod4_loop
 761 .align  32
 762 .Lmod4_loop:
 763         pclmulqdq       \$0x00,$Hkey4,$Xi
 764         xorps           $Xm,$Xmn
 765          movdqu         0x30($inp),$Xl
 766          pshufb         $T3,$Xl
 767         pclmulqdq       \$0x11,$Hkey4,$Xhi
 768         xorps           $Xln,$Xi
 769          movdqu         0x20($inp),$Xln
 770          movdqa         $Xl,$Xh
 771         pclmulqdq       \$0x10,$HK,$T1
 772          pshufd         \$0b01001110,$Xl,$Xm
 773         xorps           $Xhn,$Xhi
 774          pxor           $Xl,$Xm
 775          pshufb         $T3,$Xln
 776         movups          0x20($Htbl),$HK
 777         xorps           $Xmn,$T1
 778          pclmulqdq      \$0x00,$Hkey,$Xl
 779          pshufd         \$0b01001110,$Xln,$Xmn
 780
 781         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 782          movdqa         $Xln,$Xhn
 783         pxor            $Xhi,$T1                #
 784          pxor           $Xln,$Xmn
 785         movdqa          $T1,$T2                 #
 786          pclmulqdq      \$0x11,$Hkey,$Xh
 787         pslldq          \$8,$T1
 788         psrldq          \$8,$T2                 #
 789         pxor            $T1,$Xi
 790         movdqa          .L7_mask(%rip),$T1
 791         pxor            $T2,$Xhi                #
 792         movq            %rax,$T2
 793
 794         pand            $Xi,$T1                 # 1st phase
 795         pshufb          $T1,$T2                 #
 796         pxor            $Xi,$T2                 #
 797          pclmulqdq      \$0x00,$HK,$Xm
 798         psllq           \$57,$T2                #
 799         movdqa          $T2,$T1                 #
 800         pslldq          \$8,$T2
 801          pclmulqdq      \$0x00,$Hkey2,$Xln
 802         psrldq          \$8,$T1                 #
 803         pxor            $T2,$Xi
 804         pxor            $T1,$Xhi                #
 805         movdqu          0($inp),$T1
 806
 807         movdqa          $Xi,$T2                 # 2nd phase
 808         psrlq           \$1,$Xi
 809          pclmulqdq      \$0x11,$Hkey2,$Xhn
 810          xorps          $Xl,$Xln
 811          movdqu         0x10($inp),$Xl
 812          pshufb         $T3,$Xl
 813          pclmulqdq      \$0x10,$HK,$Xmn
 814          xorps          $Xh,$Xhn
 815          movups         0x50($Htbl),$HK
 816         pshufb          $T3,$T1
 817         pxor            $T2,$Xhi                #
 818         pxor            $Xi,$T2
 819         psrlq           \$5,$Xi
 820
 821          movdqa         $Xl,$Xh
 822          pxor           $Xm,$Xmn
 823          pshufd         \$0b01001110,$Xl,$Xm
 824         pxor            $T2,$Xi                 #
 825         pxor            $T1,$Xhi
 826          pxor           $Xl,$Xm
 827          pclmulqdq      \$0x00,$Hkey3,$Xl
 828         psrlq           \$1,$Xi                 #
 829         pxor            $Xhi,$Xi                #
 830         movdqa          $Xi,$Xhi
 831          pclmulqdq      \$0x11,$Hkey3,$Xh
 832          xorps          $Xl,$Xln
 833         pshufd          \$0b01001110,$Xi,$T1
 834         pxor            $Xi,$T1
 835
 836          pclmulqdq      \$0x00,$HK,$Xm
 837          xorps          $Xh,$Xhn
 838
 839         lea     0x40($inp),$inp
 840         sub     \$0x40,$len
 841         jnc     .Lmod4_loop
 842
 843 .Ltail4x:
 844         pclmulqdq       \$0x00,$Hkey4,$Xi
 845         pclmulqdq       \$0x11,$Hkey4,$Xhi
 846         pclmulqdq       \$0x10,$HK,$T1
 847         xorps           $Xm,$Xmn
 848         xorps           $Xln,$Xi
 849         xorps           $Xhn,$Xhi
 850         pxor            $Xi,$Xhi                # aggregated Karatsuba post-processing
 851         pxor            $Xmn,$T1
 852
 853         pxor            $Xhi,$T1                #
 854         pxor            $Xi,$Xhi
 855
 856         movdqa          $T1,$T2                 #
 857         psrldq          \$8,$T1
 858         pslldq          \$8,$T2                 #
 859         pxor            $T1,$Xhi
 860         pxor            $T2,$Xi                 #
 861 ___
 862         &reduction_alg9($Xhi,$Xi);
 863 $code.=<<___;
 864         add     \$0x40,$len
 865         jz      .Ldone
 866         movdqu  0x20($Htbl),$HK
 867         sub     \$0x10,$len
 868         jz      .Lodd_tail
 869 .Lskip4x:
 870 ___
 871 }
 872 $code.=<<___;
 873         #######
 874         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
 875         #       [(H*Ii+1) + (H*Xi+1)] mod P =
 876         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
 877         #
 878         movdqu          ($inp),$T1              # Ii
 879         movdqu          16($inp),$Xln           # Ii+1
 880         pshufb          $T3,$T1
 881         pshufb          $T3,$Xln
 882         pxor            $T1,$Xi                 # Ii+Xi
 883
 884         movdqa          $Xln,$Xhn
 885         pshufd          \$0b01001110,$Xln,$Xmn
 886         pxor            $Xln,$Xmn
 887         pclmulqdq       \$0x00,$Hkey,$Xln
 888         pclmulqdq       \$0x11,$Hkey,$Xhn
 889         pclmulqdq       \$0x00,$HK,$Xmn
 890
 891         lea             32($inp),$inp           # i+=2
 892         nop
 893         sub             \$0x20,$len
 894         jbe             .Leven_tail
 895         nop
 896         jmp             .Lmod_loop
 897
 898 .align  32
 899 .Lmod_loop:
 900         movdqa          $Xi,$Xhi
 901         movdqa          $Xmn,$T1
 902         pshufd          \$0b01001110,$Xi,$Xmn   #
 903         pxor            $Xi,$Xmn                #
 904
 905         pclmulqdq       \$0x00,$Hkey2,$Xi
 906         pclmulqdq       \$0x11,$Hkey2,$Xhi
 907         pclmulqdq       \$0x10,$HK,$Xmn
 908
 909         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 910         pxor            $Xhn,$Xhi
 911           movdqu        ($inp),$T2              # Ii
 912         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 913           pshufb        $T3,$T2
 914           movdqu        16($inp),$Xln           # Ii+1
 915
 916         pxor            $Xhi,$T1
 917           pxor          $T2,$Xhi                # "Ii+Xi", consume early
 918         pxor            $T1,$Xmn
 919          pshufb         $T3,$Xln
 920         movdqa          $Xmn,$T1                #
 921         psrldq          \$8,$T1
 922         pslldq          \$8,$Xmn                #
 923         pxor            $T1,$Xhi
 924         pxor            $Xmn,$Xi                #
 925
 926         movdqa          $Xln,$Xhn               #
 927
 928           movdqa        $Xi,$T2                 # 1st phase
 929           movdqa        $Xi,$T1
 930           psllq         \$5,$Xi
 931           pxor          $Xi,$T1                 #
 932         pclmulqdq       \$0x00,$Hkey,$Xln       #######
 933           psllq         \$1,$Xi
 934           pxor          $T1,$Xi                 #
 935           psllq         \$57,$Xi                #
 936           movdqa        $Xi,$T1                 #
 937           pslldq        \$8,$Xi
 938           psrldq        \$8,$T1                 #
 939           pxor          $T2,$Xi
 940         pshufd          \$0b01001110,$Xhn,$Xmn
 941           pxor          $T1,$Xhi                #
 942         pxor            $Xhn,$Xmn               #
 943
 944           movdqa        $Xi,$T2                 # 2nd phase
 945           psrlq         \$1,$Xi
 946         pclmulqdq       \$0x11,$Hkey,$Xhn       #######
 947           pxor          $T2,$Xhi                #
 948           pxor          $Xi,$T2
 949           psrlq         \$5,$Xi
 950           pxor          $T2,$Xi                 #
 951         lea             32($inp),$inp
 952           psrlq         \$1,$Xi                 #
 953         pclmulqdq       \$0x00,$HK,$Xmn         #######
 954           pxor          $Xhi,$Xi                #
 955
 956         sub             \$0x20,$len
 957         ja              .Lmod_loop
 958
 959 .Leven_tail:
 960          movdqa         $Xi,$Xhi
 961          movdqa         $Xmn,$T1
 962          pshufd         \$0b01001110,$Xi,$Xmn   #
 963          pxor           $Xi,$Xmn                #
 964
 965         pclmulqdq       \$0x00,$Hkey2,$Xi
 966         pclmulqdq       \$0x11,$Hkey2,$Xhi
 967         pclmulqdq       \$0x10,$HK,$Xmn
 968
 969         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 970         pxor            $Xhn,$Xhi
 971         pxor            $Xi,$T1
 972         pxor            $Xhi,$T1
 973         pxor            $T1,$Xmn
 974         movdqa          $Xmn,$T1                #
 975         psrldq          \$8,$T1
 976         pslldq          \$8,$Xmn                #
 977         pxor            $T1,$Xhi
 978         pxor            $Xmn,$Xi                #
 979 ___
 980         &reduction_alg9 ($Xhi,$Xi);
 981 $code.=<<___;
 982         test            $len,$len
 983         jnz             .Ldone
 984
 985 .Lodd_tail:
 986         movdqu          ($inp),$T1              # Ii
 987         pshufb          $T3,$T1
 988         pxor            $T1,$Xi                 # Ii+Xi
 989 ___
 990         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H*(Ii+Xi)
 991         &reduction_alg9 ($Xhi,$Xi);
 992 $code.=<<___;
 993 .Ldone:
 994         pshufb          $T3,$Xi
 995         movdqu          $Xi,($Xip)
 996 ___
 997 $code.=<<___ if ($win64);
 998         movaps  (%rsp),%xmm6
 999         movaps  0x10(%rsp),%xmm7
1000         movaps  0x20(%rsp),%xmm8
1001         movaps  0x30(%rsp),%xmm9
1002         movaps  0x40(%rsp),%xmm10
1003         movaps  0x50(%rsp),%xmm11
1004         movaps  0x60(%rsp),%xmm12
1005         movaps  0x70(%rsp),%xmm13
1006         movaps  0x80(%rsp),%xmm14
1007         movaps  0x90(%rsp),%xmm15
1008         lea     0xa8(%rsp),%rsp
1009 .LSEH_end_gcm_ghash_clmul:
1010 ___
1011 $code.=<<___;
1012         ret
1013 .cfi_endproc
1014 .size   gcm_ghash_clmul,.-gcm_ghash_clmul
1015 ___
1016 }
1017 \f
1018 $code.=<<___;
1019 .globl  gcm_init_avx
1020 .type   gcm_init_avx,\@abi-omnipotent
1021 .align  32
1022 gcm_init_avx:
1023 .cfi_startproc
1024 ___
1025 if ($avx) {
1026 my ($Htbl,$Xip)=@_4args;
1027 my $HK="%xmm6";
1028
1029 $code.=<<___ if ($win64);
1030 .LSEH_begin_gcm_init_avx:
1031         # I can't trust assembler to use specific encoding:-(
1032         .byte   0x48,0x83,0xec,0x18             #sub    $0x18,%rsp
1033         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
1034 ___
1035 $code.=<<___;
1036         vzeroupper
1037
1038         vmovdqu         ($Xip),$Hkey
1039         vpshufd         \$0b01001110,$Hkey,$Hkey        # dword swap
1040
1041         # <<1 twist
1042         vpshufd         \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
1043         vpsrlq          \$63,$Hkey,$T1
1044         vpsllq          \$1,$Hkey,$Hkey
1045         vpxor           $T3,$T3,$T3             #
1046         vpcmpgtd        $T2,$T3,$T3             # broadcast carry bit
1047         vpslldq         \$8,$T1,$T1
1048         vpor            $T1,$Hkey,$Hkey         # H<<=1
1049
1050         # magic reduction
1051         vpand           .L0x1c2_polynomial(%rip),$T3,$T3
1052         vpxor           $T3,$Hkey,$Hkey         # if(carry) H^=0x1c2_polynomial
1053
1054         vpunpckhqdq     $Hkey,$Hkey,$HK
1055         vmovdqa         $Hkey,$Xi
1056         vpxor           $Hkey,$HK,$HK
1057         mov             \$4,%r10                # up to H^8
1058         jmp             .Linit_start_avx
1059 ___
1060
1061 sub clmul64x64_avx {
1062 my ($Xhi,$Xi,$Hkey,$HK)=@_;
1063
1064 if (!defined($HK)) {    $HK = $T2;
1065 $code.=<<___;
1066         vpunpckhqdq     $Xi,$Xi,$T1
1067         vpunpckhqdq     $Hkey,$Hkey,$T2
1068         vpxor           $Xi,$T1,$T1             #
1069         vpxor           $Hkey,$T2,$T2
1070 ___
1071 } else {
1072 $code.=<<___;
1073         vpunpckhqdq     $Xi,$Xi,$T1
1074         vpxor           $Xi,$T1,$T1             #
1075 ___
1076 }
1077 $code.=<<___;
1078         vpclmulqdq      \$0x11,$Hkey,$Xi,$Xhi   #######
1079         vpclmulqdq      \$0x00,$Hkey,$Xi,$Xi    #######
1080         vpclmulqdq      \$0x00,$HK,$T1,$T1      #######
1081         vpxor           $Xi,$Xhi,$T2            #
1082         vpxor           $T2,$T1,$T1             #
1083
1084         vpslldq         \$8,$T1,$T2             #
1085         vpsrldq         \$8,$T1,$T1
1086         vpxor           $T2,$Xi,$Xi             #
1087         vpxor           $T1,$Xhi,$Xhi
1088 ___
1089 }
1090
1091 sub reduction_avx {
1092 my ($Xhi,$Xi) = @_;
1093
1094 $code.=<<___;
1095         vpsllq          \$57,$Xi,$T1            # 1st phase
1096         vpsllq          \$62,$Xi,$T2
1097         vpxor           $T1,$T2,$T2             #
1098         vpsllq          \$63,$Xi,$T1
1099         vpxor           $T1,$T2,$T2             #
1100         vpslldq         \$8,$T2,$T1             #
1101         vpsrldq         \$8,$T2,$T2
1102         vpxor           $T1,$Xi,$Xi             #
1103         vpxor           $T2,$Xhi,$Xhi
1104
1105         vpsrlq          \$1,$Xi,$T2             # 2nd phase
1106         vpxor           $Xi,$Xhi,$Xhi
1107         vpxor           $T2,$Xi,$Xi             #
1108         vpsrlq          \$5,$T2,$T2
1109         vpxor           $T2,$Xi,$Xi             #
1110         vpsrlq          \$1,$Xi,$Xi             #
1111         vpxor           $Xhi,$Xi,$Xi            #
1112 ___
1113 }
1114
1115 $code.=<<___;
1116 .align  32
1117 .Linit_loop_avx:
1118         vpalignr        \$8,$T1,$T2,$T3         # low part is H.lo^H.hi...
1119         vmovdqu         $T3,-0x10($Htbl)        # save Karatsuba "salt"
1120 ___
1121         &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK);   # calculate H^3,5,7
1122         &reduction_avx  ($Xhi,$Xi);
1123 $code.=<<___;
1124 .Linit_start_avx:
1125         vmovdqa         $Xi,$T3
1126 ___
1127         &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK);   # calculate H^2,4,6,8
1128         &reduction_avx  ($Xhi,$Xi);
1129 $code.=<<___;
1130         vpshufd         \$0b01001110,$T3,$T1
1131         vpshufd         \$0b01001110,$Xi,$T2
1132         vpxor           $T3,$T1,$T1             # Karatsuba pre-processing
1133         vmovdqu         $T3,0x00($Htbl)         # save H^1,3,5,7
1134         vpxor           $Xi,$T2,$T2             # Karatsuba pre-processing
1135         vmovdqu         $Xi,0x10($Htbl)         # save H^2,4,6,8
1136         lea             0x30($Htbl),$Htbl
1137         sub             \$1,%r10
1138         jnz             .Linit_loop_avx
1139
1140         vpalignr        \$8,$T2,$T1,$T3         # last "salt" is flipped
1141         vmovdqu         $T3,-0x10($Htbl)
1142
1143         vzeroupper
1144 ___
1145 $code.=<<___ if ($win64);
1146         movaps  (%rsp),%xmm6
1147         lea     0x18(%rsp),%rsp
1148 .LSEH_end_gcm_init_avx:
1149 ___
1150 $code.=<<___;
1151         ret
1152 .cfi_endproc
1153 .size   gcm_init_avx,.-gcm_init_avx
1154 ___
1155 } else {
1156 $code.=<<___;
1157         jmp     .L_init_clmul
1158 .cfi_endproc
1159 .size   gcm_init_avx,.-gcm_init_avx
1160 ___
1161 }
1162
1163 $code.=<<___;
1164 .globl  gcm_gmult_avx
1165 .type   gcm_gmult_avx,\@abi-omnipotent
1166 .align  32
1167 gcm_gmult_avx:
1168 .cfi_startproc
1169         jmp     .L_gmult_clmul
1170 .cfi_endproc
1171 .size   gcm_gmult_avx,.-gcm_gmult_avx
1172 ___
1173 \f
1174 $code.=<<___;
1175 .globl  gcm_ghash_avx
1176 .type   gcm_ghash_avx,\@abi-omnipotent
1177 .align  32
1178 gcm_ghash_avx:
1179 .cfi_startproc
1180 ___
1181 if ($avx) {
1182 my ($Xip,$Htbl,$inp,$len)=@_4args;
1183 my ($Xlo,$Xhi,$Xmi,
1184     $Zlo,$Zhi,$Zmi,
1185     $Hkey,$HK,$T1,$T2,
1186     $Xi,$Xo,$Tred,$bswap,$Ii,$Ij) = map("%xmm$_",(0..15));
1187
1188 $code.=<<___ if ($win64);
1189         lea     -0x88(%rsp),%rax
1190 .LSEH_begin_gcm_ghash_avx:
1191         # I can't trust assembler to use specific encoding:-(
1192         .byte   0x48,0x8d,0x60,0xe0             #lea    -0x20(%rax),%rsp
1193         .byte   0x0f,0x29,0x70,0xe0             #movaps %xmm6,-0x20(%rax)
1194         .byte   0x0f,0x29,0x78,0xf0             #movaps %xmm7,-0x10(%rax)
1195         .byte   0x44,0x0f,0x29,0x00             #movaps %xmm8,0(%rax)
1196         .byte   0x44,0x0f,0x29,0x48,0x10        #movaps %xmm9,0x10(%rax)
1197         .byte   0x44,0x0f,0x29,0x50,0x20        #movaps %xmm10,0x20(%rax)
1198         .byte   0x44,0x0f,0x29,0x58,0x30        #movaps %xmm11,0x30(%rax)
1199         .byte   0x44,0x0f,0x29,0x60,0x40        #movaps %xmm12,0x40(%rax)
1200         .byte   0x44,0x0f,0x29,0x68,0x50        #movaps %xmm13,0x50(%rax)
1201         .byte   0x44,0x0f,0x29,0x70,0x60        #movaps %xmm14,0x60(%rax)
1202         .byte   0x44,0x0f,0x29,0x78,0x70        #movaps %xmm15,0x70(%rax)
1203 ___
1204 $code.=<<___;
1205         vzeroupper
1206
1207         vmovdqu         ($Xip),$Xi              # load $Xi
1208         lea             .L0x1c2_polynomial(%rip),%r10
1209         lea             0x40($Htbl),$Htbl       # size optimization
1210         vmovdqu         .Lbswap_mask(%rip),$bswap
1211         vpshufb         $bswap,$Xi,$Xi
1212         cmp             \$0x80,$len
1213         jb              .Lshort_avx
1214         sub             \$0x80,$len
1215
1216         vmovdqu         0x70($inp),$Ii          # I[7]
1217         vmovdqu         0x00-0x40($Htbl),$Hkey  # $Hkey^1
1218         vpshufb         $bswap,$Ii,$Ii
1219         vmovdqu         0x20-0x40($Htbl),$HK
1220
1221         vpunpckhqdq     $Ii,$Ii,$T2
1222          vmovdqu        0x60($inp),$Ij          # I[6]
1223         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1224         vpxor           $Ii,$T2,$T2
1225          vpshufb        $bswap,$Ij,$Ij
1226         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1227          vmovdqu        0x10-0x40($Htbl),$Hkey  # $Hkey^2
1228          vpunpckhqdq    $Ij,$Ij,$T1
1229          vmovdqu        0x50($inp),$Ii          # I[5]
1230         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1231          vpxor          $Ij,$T1,$T1
1232
1233          vpshufb        $bswap,$Ii,$Ii
1234         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1235          vpunpckhqdq    $Ii,$Ii,$T2
1236         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1237          vmovdqu        0x30-0x40($Htbl),$Hkey  # $Hkey^3
1238          vpxor          $Ii,$T2,$T2
1239          vmovdqu        0x40($inp),$Ij          # I[4]
1240         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1241          vmovdqu        0x50-0x40($Htbl),$HK
1242
1243          vpshufb        $bswap,$Ij,$Ij
1244         vpxor           $Xlo,$Zlo,$Zlo
1245         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1246         vpxor           $Xhi,$Zhi,$Zhi
1247          vpunpckhqdq    $Ij,$Ij,$T1
1248         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1249          vmovdqu        0x40-0x40($Htbl),$Hkey  # $Hkey^4
1250         vpxor           $Xmi,$Zmi,$Zmi
1251         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1252          vpxor          $Ij,$T1,$T1
1253
1254          vmovdqu        0x30($inp),$Ii          # I[3]
1255         vpxor           $Zlo,$Xlo,$Xlo
1256         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1257         vpxor           $Zhi,$Xhi,$Xhi
1258          vpshufb        $bswap,$Ii,$Ii
1259         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1260          vmovdqu        0x60-0x40($Htbl),$Hkey  # $Hkey^5
1261         vpxor           $Zmi,$Xmi,$Xmi
1262          vpunpckhqdq    $Ii,$Ii,$T2
1263         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1264          vmovdqu        0x80-0x40($Htbl),$HK
1265          vpxor          $Ii,$T2,$T2
1266
1267          vmovdqu        0x20($inp),$Ij          # I[2]
1268         vpxor           $Xlo,$Zlo,$Zlo
1269         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1270         vpxor           $Xhi,$Zhi,$Zhi
1271          vpshufb        $bswap,$Ij,$Ij
1272         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1273          vmovdqu        0x70-0x40($Htbl),$Hkey  # $Hkey^6
1274         vpxor           $Xmi,$Zmi,$Zmi
1275          vpunpckhqdq    $Ij,$Ij,$T1
1276         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1277          vpxor          $Ij,$T1,$T1
1278
1279          vmovdqu        0x10($inp),$Ii          # I[1]
1280         vpxor           $Zlo,$Xlo,$Xlo
1281         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1282         vpxor           $Zhi,$Xhi,$Xhi
1283          vpshufb        $bswap,$Ii,$Ii
1284         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1285          vmovdqu        0x90-0x40($Htbl),$Hkey  # $Hkey^7
1286         vpxor           $Zmi,$Xmi,$Xmi
1287          vpunpckhqdq    $Ii,$Ii,$T2
1288         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1289          vmovdqu        0xb0-0x40($Htbl),$HK
1290          vpxor          $Ii,$T2,$T2
1291
1292          vmovdqu        ($inp),$Ij              # I[0]
1293         vpxor           $Xlo,$Zlo,$Zlo
1294         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1295         vpxor           $Xhi,$Zhi,$Zhi
1296          vpshufb        $bswap,$Ij,$Ij
1297         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1298          vmovdqu        0xa0-0x40($Htbl),$Hkey  # $Hkey^8
1299         vpxor           $Xmi,$Zmi,$Zmi
1300         vpclmulqdq      \$0x10,$HK,$T2,$Xmi
1301
1302         lea             0x80($inp),$inp
1303         cmp             \$0x80,$len
1304         jb              .Ltail_avx
1305
1306         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1307         sub             \$0x80,$len
1308         jmp             .Loop8x_avx
1309
1310 .align  32
1311 .Loop8x_avx:
1312         vpunpckhqdq     $Ij,$Ij,$T1
1313          vmovdqu        0x70($inp),$Ii          # I[7]
1314         vpxor           $Xlo,$Zlo,$Zlo
1315         vpxor           $Ij,$T1,$T1
1316         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xi
1317          vpshufb        $bswap,$Ii,$Ii
1318         vpxor           $Xhi,$Zhi,$Zhi
1319         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xo
1320          vmovdqu        0x00-0x40($Htbl),$Hkey  # $Hkey^1
1321          vpunpckhqdq    $Ii,$Ii,$T2
1322         vpxor           $Xmi,$Zmi,$Zmi
1323         vpclmulqdq      \$0x00,$HK,$T1,$Tred
1324          vmovdqu        0x20-0x40($Htbl),$HK
1325          vpxor          $Ii,$T2,$T2
1326
1327           vmovdqu       0x60($inp),$Ij          # I[6]
1328          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1329         vpxor           $Zlo,$Xi,$Xi            # collect result
1330           vpshufb       $bswap,$Ij,$Ij
1331          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1332         vxorps          $Zhi,$Xo,$Xo
1333           vmovdqu       0x10-0x40($Htbl),$Hkey  # $Hkey^2
1334          vpunpckhqdq    $Ij,$Ij,$T1
1335          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1336         vpxor           $Zmi,$Tred,$Tred
1337          vxorps         $Ij,$T1,$T1
1338
1339           vmovdqu       0x50($inp),$Ii          # I[5]
1340         vpxor           $Xi,$Tred,$Tred         # aggregated Karatsuba post-processing
1341          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1342         vpxor           $Xo,$Tred,$Tred
1343         vpslldq         \$8,$Tred,$T2
1344          vpxor          $Xlo,$Zlo,$Zlo
1345          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1346         vpsrldq         \$8,$Tred,$Tred
1347         vpxor           $T2, $Xi, $Xi
1348           vmovdqu       0x30-0x40($Htbl),$Hkey  # $Hkey^3
1349           vpshufb       $bswap,$Ii,$Ii
1350         vxorps          $Tred,$Xo, $Xo
1351          vpxor          $Xhi,$Zhi,$Zhi
1352          vpunpckhqdq    $Ii,$Ii,$T2
1353          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1354           vmovdqu       0x50-0x40($Htbl),$HK
1355          vpxor          $Ii,$T2,$T2
1356          vpxor          $Xmi,$Zmi,$Zmi
1357
1358           vmovdqu       0x40($inp),$Ij          # I[4]
1359         vpalignr        \$8,$Xi,$Xi,$Tred       # 1st phase
1360          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1361           vpshufb       $bswap,$Ij,$Ij
1362          vpxor          $Zlo,$Xlo,$Xlo
1363          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1364           vmovdqu       0x40-0x40($Htbl),$Hkey  # $Hkey^4
1365          vpunpckhqdq    $Ij,$Ij,$T1
1366          vpxor          $Zhi,$Xhi,$Xhi
1367          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1368          vxorps         $Ij,$T1,$T1
1369          vpxor          $Zmi,$Xmi,$Xmi
1370
1371           vmovdqu       0x30($inp),$Ii          # I[3]
1372         vpclmulqdq      \$0x10,(%r10),$Xi,$Xi
1373          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1374           vpshufb       $bswap,$Ii,$Ii
1375          vpxor          $Xlo,$Zlo,$Zlo
1376          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1377           vmovdqu       0x60-0x40($Htbl),$Hkey  # $Hkey^5
1378          vpunpckhqdq    $Ii,$Ii,$T2
1379          vpxor          $Xhi,$Zhi,$Zhi
1380          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1381           vmovdqu       0x80-0x40($Htbl),$HK
1382          vpxor          $Ii,$T2,$T2
1383          vpxor          $Xmi,$Zmi,$Zmi
1384
1385           vmovdqu       0x20($inp),$Ij          # I[2]
1386          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1387           vpshufb       $bswap,$Ij,$Ij
1388          vpxor          $Zlo,$Xlo,$Xlo
1389          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1390           vmovdqu       0x70-0x40($Htbl),$Hkey  # $Hkey^6
1391          vpunpckhqdq    $Ij,$Ij,$T1
1392          vpxor          $Zhi,$Xhi,$Xhi
1393          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1394          vpxor          $Ij,$T1,$T1
1395          vpxor          $Zmi,$Xmi,$Xmi
1396         vxorps          $Tred,$Xi,$Xi
1397
1398           vmovdqu       0x10($inp),$Ii          # I[1]
1399         vpalignr        \$8,$Xi,$Xi,$Tred       # 2nd phase
1400          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1401           vpshufb       $bswap,$Ii,$Ii
1402          vpxor          $Xlo,$Zlo,$Zlo
1403          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1404           vmovdqu       0x90-0x40($Htbl),$Hkey  # $Hkey^7
1405         vpclmulqdq      \$0x10,(%r10),$Xi,$Xi
1406         vxorps          $Xo,$Tred,$Tred
1407          vpunpckhqdq    $Ii,$Ii,$T2
1408          vpxor          $Xhi,$Zhi,$Zhi
1409          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1410           vmovdqu       0xb0-0x40($Htbl),$HK
1411          vpxor          $Ii,$T2,$T2
1412          vpxor          $Xmi,$Zmi,$Zmi
1413
1414           vmovdqu       ($inp),$Ij              # I[0]
1415          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1416           vpshufb       $bswap,$Ij,$Ij
1417          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1418           vmovdqu       0xa0-0x40($Htbl),$Hkey  # $Hkey^8
1419         vpxor           $Tred,$Ij,$Ij
1420          vpclmulqdq     \$0x10,$HK,  $T2,$Xmi
1421         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1422
1423         lea             0x80($inp),$inp
1424         sub             \$0x80,$len
1425         jnc             .Loop8x_avx
1426
1427         add             \$0x80,$len
1428         jmp             .Ltail_no_xor_avx
1429
1430 .align  32
1431 .Lshort_avx:
1432         vmovdqu         -0x10($inp,$len),$Ii    # very last word
1433         lea             ($inp,$len),$inp
1434         vmovdqu         0x00-0x40($Htbl),$Hkey  # $Hkey^1
1435         vmovdqu         0x20-0x40($Htbl),$HK
1436         vpshufb         $bswap,$Ii,$Ij
1437
1438         vmovdqa         $Xlo,$Zlo               # subtle way to zero $Zlo,
1439         vmovdqa         $Xhi,$Zhi               # $Zhi and
1440         vmovdqa         $Xmi,$Zmi               # $Zmi
1441         sub             \$0x10,$len
1442         jz              .Ltail_avx
1443
1444         vpunpckhqdq     $Ij,$Ij,$T1
1445         vpxor           $Xlo,$Zlo,$Zlo
1446         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1447         vpxor           $Ij,$T1,$T1
1448          vmovdqu        -0x20($inp),$Ii
1449         vpxor           $Xhi,$Zhi,$Zhi
1450         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1451         vmovdqu         0x10-0x40($Htbl),$Hkey  # $Hkey^2
1452          vpshufb        $bswap,$Ii,$Ij
1453         vpxor           $Xmi,$Zmi,$Zmi
1454         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1455         vpsrldq         \$8,$HK,$HK
1456         sub             \$0x10,$len
1457         jz              .Ltail_avx
1458
1459         vpunpckhqdq     $Ij,$Ij,$T1
1460         vpxor           $Xlo,$Zlo,$Zlo
1461         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1462         vpxor           $Ij,$T1,$T1
1463          vmovdqu        -0x30($inp),$Ii
1464         vpxor           $Xhi,$Zhi,$Zhi
1465         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1466         vmovdqu         0x30-0x40($Htbl),$Hkey  # $Hkey^3
1467          vpshufb        $bswap,$Ii,$Ij
1468         vpxor           $Xmi,$Zmi,$Zmi
1469         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1470         vmovdqu         0x50-0x40($Htbl),$HK
1471         sub             \$0x10,$len
1472         jz              .Ltail_avx
1473
1474         vpunpckhqdq     $Ij,$Ij,$T1
1475         vpxor           $Xlo,$Zlo,$Zlo
1476         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1477         vpxor           $Ij,$T1,$T1
1478          vmovdqu        -0x40($inp),$Ii
1479         vpxor           $Xhi,$Zhi,$Zhi
1480         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1481         vmovdqu         0x40-0x40($Htbl),$Hkey  # $Hkey^4
1482          vpshufb        $bswap,$Ii,$Ij
1483         vpxor           $Xmi,$Zmi,$Zmi
1484         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1485         vpsrldq         \$8,$HK,$HK
1486         sub             \$0x10,$len
1487         jz              .Ltail_avx
1488
1489         vpunpckhqdq     $Ij,$Ij,$T1
1490         vpxor           $Xlo,$Zlo,$Zlo
1491         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1492         vpxor           $Ij,$T1,$T1
1493          vmovdqu        -0x50($inp),$Ii
1494         vpxor           $Xhi,$Zhi,$Zhi
1495         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1496         vmovdqu         0x60-0x40($Htbl),$Hkey  # $Hkey^5
1497          vpshufb        $bswap,$Ii,$Ij
1498         vpxor           $Xmi,$Zmi,$Zmi
1499         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1500         vmovdqu         0x80-0x40($Htbl),$HK
1501         sub             \$0x10,$len
1502         jz              .Ltail_avx
1503
1504         vpunpckhqdq     $Ij,$Ij,$T1
1505         vpxor           $Xlo,$Zlo,$Zlo
1506         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1507         vpxor           $Ij,$T1,$T1
1508          vmovdqu        -0x60($inp),$Ii
1509         vpxor           $Xhi,$Zhi,$Zhi
1510         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1511         vmovdqu         0x70-0x40($Htbl),$Hkey  # $Hkey^6
1512          vpshufb        $bswap,$Ii,$Ij
1513         vpxor           $Xmi,$Zmi,$Zmi
1514         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1515         vpsrldq         \$8,$HK,$HK
1516         sub             \$0x10,$len
1517         jz              .Ltail_avx
1518
1519         vpunpckhqdq     $Ij,$Ij,$T1
1520         vpxor           $Xlo,$Zlo,$Zlo
1521         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1522         vpxor           $Ij,$T1,$T1
1523          vmovdqu        -0x70($inp),$Ii
1524         vpxor           $Xhi,$Zhi,$Zhi
1525         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1526         vmovdqu         0x90-0x40($Htbl),$Hkey  # $Hkey^7
1527          vpshufb        $bswap,$Ii,$Ij
1528         vpxor           $Xmi,$Zmi,$Zmi
1529         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1530         vmovq           0xb8-0x40($Htbl),$HK
1531         sub             \$0x10,$len
1532         jmp             .Ltail_avx
1533
1534 .align  32
1535 .Ltail_avx:
1536         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1537 .Ltail_no_xor_avx:
1538         vpunpckhqdq     $Ij,$Ij,$T1
1539         vpxor           $Xlo,$Zlo,$Zlo
1540         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1541         vpxor           $Ij,$T1,$T1
1542         vpxor           $Xhi,$Zhi,$Zhi
1543         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1544         vpxor           $Xmi,$Zmi,$Zmi
1545         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1546
1547         vmovdqu         (%r10),$Tred
1548
1549         vpxor           $Xlo,$Zlo,$Xi
1550         vpxor           $Xhi,$Zhi,$Xo
1551         vpxor           $Xmi,$Zmi,$Zmi
1552
1553         vpxor           $Xi, $Zmi,$Zmi          # aggregated Karatsuba post-processing
1554         vpxor           $Xo, $Zmi,$Zmi
1555         vpslldq         \$8, $Zmi,$T2
1556         vpsrldq         \$8, $Zmi,$Zmi
1557         vpxor           $T2, $Xi, $Xi
1558         vpxor           $Zmi,$Xo, $Xo
1559
1560         vpclmulqdq      \$0x10,$Tred,$Xi,$T2    # 1st phase
1561         vpalignr        \$8,$Xi,$Xi,$Xi
1562         vpxor           $T2,$Xi,$Xi
1563
1564         vpclmulqdq      \$0x10,$Tred,$Xi,$T2    # 2nd phase
1565         vpalignr        \$8,$Xi,$Xi,$Xi
1566         vpxor           $Xo,$Xi,$Xi
1567         vpxor           $T2,$Xi,$Xi
1568
1569         cmp             \$0,$len
1570         jne             .Lshort_avx
1571
1572         vpshufb         $bswap,$Xi,$Xi
1573         vmovdqu         $Xi,($Xip)
1574         vzeroupper
1575 ___
1576 $code.=<<___ if ($win64);
1577         movaps  (%rsp),%xmm6
1578         movaps  0x10(%rsp),%xmm7
1579         movaps  0x20(%rsp),%xmm8
1580         movaps  0x30(%rsp),%xmm9
1581         movaps  0x40(%rsp),%xmm10
1582         movaps  0x50(%rsp),%xmm11
1583         movaps  0x60(%rsp),%xmm12
1584         movaps  0x70(%rsp),%xmm13
1585         movaps  0x80(%rsp),%xmm14
1586         movaps  0x90(%rsp),%xmm15
1587         lea     0xa8(%rsp),%rsp
1588 .LSEH_end_gcm_ghash_avx:
1589 ___
1590 $code.=<<___;
1591         ret
1592 .cfi_endproc
1593 .size   gcm_ghash_avx,.-gcm_ghash_avx
1594 ___
1595 } else {
1596 $code.=<<___;
1597         jmp     .L_ghash_clmul
1598 .cfi_endproc
1599 .size   gcm_ghash_avx,.-gcm_ghash_avx
1600 ___
1601 }
1602 \f
1603 $code.=<<___;
1604 .align  64
1605 .Lbswap_mask:
1606         .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
1607 .L0x1c2_polynomial:
1608         .byte   1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
1609 .L7_mask:
1610         .long   7,0,7,0
1611 .L7_mask_poly:
1612         .long   7,0,`0xE1<<1`,0
1613 .align  64
1614 .type   .Lrem_4bit,\@object
1615 .Lrem_4bit:
1616         .long   0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
1617         .long   0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
1618         .long   0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
1619         .long   0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
1620 .type   .Lrem_8bit,\@object
1621 .Lrem_8bit:
1622         .value  0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
1623         .value  0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
1624         .value  0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
1625         .value  0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
1626         .value  0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
1627         .value  0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
1628         .value  0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
1629         .value  0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
1630         .value  0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
1631         .value  0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
1632         .value  0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
1633         .value  0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
1634         .value  0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
1635         .value  0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
1636         .value  0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
1637         .value  0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
1638         .value  0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
1639         .value  0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
1640         .value  0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
1641         .value  0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
1642         .value  0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
1643         .value  0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
1644         .value  0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
1645         .value  0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
1646         .value  0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
1647         .value  0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
1648         .value  0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
1649         .value  0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
1650         .value  0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
1651         .value  0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
1652         .value  0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
1653         .value  0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
1654
1655 .asciz  "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1656 .align  64
1657 ___
1658 \f
1659 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1660 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
1661 if ($win64) {
1662 $rec="%rcx";
1663 $frame="%rdx";
1664 $context="%r8";
1665 $disp="%r9";
1666
1667 $code.=<<___;
1668 .extern __imp_RtlVirtualUnwind
1669 .type   se_handler,\@abi-omnipotent
1670 .align  16
1671 se_handler:
1672         push    %rsi
1673         push    %rdi
1674         push    %rbx
1675         push    %rbp
1676         push    %r12
1677         push    %r13
1678         push    %r14
1679         push    %r15
1680         pushfq
1681         sub     \$64,%rsp
1682
1683         mov     120($context),%rax      # pull context->Rax
1684         mov     248($context),%rbx      # pull context->Rip
1685
1686         mov     8($disp),%rsi           # disp->ImageBase
1687         mov     56($disp),%r11          # disp->HandlerData
1688
1689         mov     0(%r11),%r10d           # HandlerData[0]
1690         lea     (%rsi,%r10),%r10        # prologue label
1691         cmp     %r10,%rbx               # context->Rip<prologue label
1692         jb      .Lin_prologue
1693
1694         mov     152($context),%rax      # pull context->Rsp
1695
1696         mov     4(%r11),%r10d           # HandlerData[1]
1697         lea     (%rsi,%r10),%r10        # epilogue label
1698         cmp     %r10,%rbx               # context->Rip>=epilogue label
1699         jae     .Lin_prologue
1700
1701         lea     48+280(%rax),%rax       # adjust "rsp"
1702
1703         mov     -8(%rax),%rbx
1704         mov     -16(%rax),%rbp
1705         mov     -24(%rax),%r12
1706         mov     -32(%rax),%r13
1707         mov     -40(%rax),%r14
1708         mov     -48(%rax),%r15
1709         mov     %rbx,144($context)      # restore context->Rbx
1710         mov     %rbp,160($context)      # restore context->Rbp
1711         mov     %r12,216($context)      # restore context->R12
1712         mov     %r13,224($context)      # restore context->R13
1713         mov     %r14,232($context)      # restore context->R14
1714         mov     %r15,240($context)      # restore context->R15
1715
1716 .Lin_prologue:
1717         mov     8(%rax),%rdi
1718         mov     16(%rax),%rsi
1719         mov     %rax,152($context)      # restore context->Rsp
1720         mov     %rsi,168($context)      # restore context->Rsi
1721         mov     %rdi,176($context)      # restore context->Rdi
1722
1723         mov     40($disp),%rdi          # disp->ContextRecord
1724         mov     $context,%rsi           # context
1725         mov     \$`1232/8`,%ecx         # sizeof(CONTEXT)
1726         .long   0xa548f3fc              # cld; rep movsq
1727
1728         mov     $disp,%rsi
1729         xor     %rcx,%rcx               # arg1, UNW_FLAG_NHANDLER
1730         mov     8(%rsi),%rdx            # arg2, disp->ImageBase
1731         mov     0(%rsi),%r8             # arg3, disp->ControlPc
1732         mov     16(%rsi),%r9            # arg4, disp->FunctionEntry
1733         mov     40(%rsi),%r10           # disp->ContextRecord
1734         lea     56(%rsi),%r11           # &disp->HandlerData
1735         lea     24(%rsi),%r12           # &disp->EstablisherFrame
1736         mov     %r10,32(%rsp)           # arg5
1737         mov     %r11,40(%rsp)           # arg6
1738         mov     %r12,48(%rsp)           # arg7
1739         mov     %rcx,56(%rsp)           # arg8, (NULL)
1740         call    *__imp_RtlVirtualUnwind(%rip)
1741
1742         mov     \$1,%eax                # ExceptionContinueSearch
1743         add     \$64,%rsp
1744         popfq
1745         pop     %r15
1746         pop     %r14
1747         pop     %r13
1748         pop     %r12
1749         pop     %rbp
1750         pop     %rbx
1751         pop     %rdi
1752         pop     %rsi
1753         ret
1754 .size   se_handler,.-se_handler
1755
1756 .section        .pdata
1757 .align  4
1758         .rva    .LSEH_begin_gcm_gmult_4bit
1759         .rva    .LSEH_end_gcm_gmult_4bit
1760         .rva    .LSEH_info_gcm_gmult_4bit
1761
1762         .rva    .LSEH_begin_gcm_ghash_4bit
1763         .rva    .LSEH_end_gcm_ghash_4bit
1764         .rva    .LSEH_info_gcm_ghash_4bit
1765
1766         .rva    .LSEH_begin_gcm_init_clmul
1767         .rva    .LSEH_end_gcm_init_clmul
1768         .rva    .LSEH_info_gcm_init_clmul
1769
1770         .rva    .LSEH_begin_gcm_ghash_clmul
1771         .rva    .LSEH_end_gcm_ghash_clmul
1772         .rva    .LSEH_info_gcm_ghash_clmul
1773 ___
1774 $code.=<<___    if ($avx);
1775         .rva    .LSEH_begin_gcm_init_avx
1776         .rva    .LSEH_end_gcm_init_avx
1777         .rva    .LSEH_info_gcm_init_clmul
1778
1779         .rva    .LSEH_begin_gcm_ghash_avx
1780         .rva    .LSEH_end_gcm_ghash_avx
1781         .rva    .LSEH_info_gcm_ghash_clmul
1782 ___
1783 $code.=<<___;
1784 .section        .xdata
1785 .align  8
1786 .LSEH_info_gcm_gmult_4bit:
1787         .byte   9,0,0,0
1788         .rva    se_handler
1789         .rva    .Lgmult_prologue,.Lgmult_epilogue       # HandlerData
1790 .LSEH_info_gcm_ghash_4bit:
1791         .byte   9,0,0,0
1792         .rva    se_handler
1793         .rva    .Lghash_prologue,.Lghash_epilogue       # HandlerData
1794 .LSEH_info_gcm_init_clmul:
1795         .byte   0x01,0x08,0x03,0x00
1796         .byte   0x08,0x68,0x00,0x00     #movaps 0x00(rsp),xmm6
1797         .byte   0x04,0x22,0x00,0x00     #sub    rsp,0x18
1798 .LSEH_info_gcm_ghash_clmul:
1799         .byte   0x01,0x33,0x16,0x00
1800         .byte   0x33,0xf8,0x09,0x00     #movaps 0x90(rsp),xmm15
1801         .byte   0x2e,0xe8,0x08,0x00     #movaps 0x80(rsp),xmm14
1802         .byte   0x29,0xd8,0x07,0x00     #movaps 0x70(rsp),xmm13
1803         .byte   0x24,0xc8,0x06,0x00     #movaps 0x60(rsp),xmm12
1804         .byte   0x1f,0xb8,0x05,0x00     #movaps 0x50(rsp),xmm11
1805         .byte   0x1a,0xa8,0x04,0x00     #movaps 0x40(rsp),xmm10
1806         .byte   0x15,0x98,0x03,0x00     #movaps 0x30(rsp),xmm9
1807         .byte   0x10,0x88,0x02,0x00     #movaps 0x20(rsp),xmm8
1808         .byte   0x0c,0x78,0x01,0x00     #movaps 0x10(rsp),xmm7
1809         .byte   0x08,0x68,0x00,0x00     #movaps 0x00(rsp),xmm6
1810         .byte   0x04,0x01,0x15,0x00     #sub    rsp,0xa8
1811 ___
1812 }
1813 \f
1814 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
1815
1816 print $code;
1817
1818 close STDOUT;